Computer Architecture, Alan Perlis

Alan J. Perlis was not only a wonderful human being, an excellent mentor to many computer scientists, and a talented coiner of aphorisms, he was also a great storyteller. I remember one of his stories concerned computer architecture. I may not have remembered all the details exactly, but here is the gist:

When we were preparing to take delivery of the Bendix G-20 computer at Carnegie [1] a man dropped into my office and said he had some "architectural questions" about the new computer. Being busy, I handed him the programming manual and told him it would answer all his questions about the architecture of the G-20.

However, he returned a couple of days later, saying he still had some unanswered questions. So I handed him the hardware manual, and said that should cover everything that wasn't in the programming manual.

He returned more quickly this time, saying "I don't know why you gave me those manuals, but if someone doesn't tell me how much this computer weighs, I won't know how much bracing the floor is going to require!"

Note:
[1] Carnegie Institute of Technology, now Carnegie Mellon University.

Bendix G-15 Architecture

The Bendix G-15D represented an architectural dead end. Only a few later commercial computers, such as the Packard Bell PB 250, followed the same lines, and none of them were as successful as the first-generation G-15 and LGP-30. Even Bendix went to an architecture much closer to Intercom than to the G-15 for their transistorized computer, the G-20 (which also cost 10 times as much).

So the G-15 architecture is interesting today primarily as an example of how optimizing for recirculating memory and vacuum-tube plus diode logic leads to very different trade-offs than for random-access memory and transistor logic.

Vacuum tubes took a lot of space (over a cubic inch apiece), used a lot of power (several watts), and were the most failure-prone class of component. A flip-flop took two tubes (but provided both the signal and its negation), logical not could be done with one tube, and boolean expressions in disjunctive normal form (ors of ands) could be done purely with germanium diodes and resistors. So evaluating complicated boolean expressions was a lot cheaper than remembering something from one "bit time" to the next.

However, I have a different motivation for discussing some of its characteristics in this blog: It will set context for some of the stories I have to tell.

• Bit-serial arithmetic.
• Memory model.
• Instruction set architecture.

Photo by The Martin Company, Baltimore

The 26 neon lights (grouped above the various meters and knobs) displayed essentially all the non-recirculating state (held in vacuum tube flip-flops) visible to the programmer, and were heavily used for debugging.

Also note the punched paper tape emerging from near the top of the machine, just above the paper tape cassette.


Photo by U. S. Army Map Service

(Both photos are links to Ed Thelen's excellent historical site.)